Erik Reinertsen     Blog     Research     Tech     Teaching     Talks

AI for medical imaging


Here I provide a highly simplified overview of machine learning, list recent papers describing the application of these approaches for medical imaging, and suggest resources for further learning.

This post is aimed at clinicians or scientists that do not have backgrounds in computer science, engineering, or statistics.

Machine learning (ML)

ML has three broad categories: supervised learning, unsupervised learning, and reinforcement learning.

  1. Supervised learning

    • The algorithm is given labeled data or ‘features’ (e.g. cystoscopy image from a patient) with paired labels (that same patient’s comorbidities, length of stay, infection, hematuria, pain, etc.))
    • The algorithm “learns” or maps the associations between the features and the labels.
    • A good algorithm can generalize in the sense that it can see new features from new patients and then correctly predict what the labels (outcomes) will be.
    • See https://www.youtube.com/watch?v=bQI5uDxrFfA for a great summary by one of the leaders in AI.
  2. Unsupervised learning

    • Evaluates hidden structure in data that do not have labels, e.g. clustering your data and seeing if the clusters have clinical meaning.
    • See https://www.youtube.com/watch?v=Ev8YbxPu_bQ
  3. Reinforcement learning

    • Framework for how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward.
    • This is partially how self-driving cars and Google’s AlphaGo works.
    • Has been used to optimize heparin dosing guidelines and maintained therapeutic INR levels.
    • Review paper of these concepts applied to medical imaging: Erickson et al. Machine Learning for Medical Imaging. RadioGraphics 2017 37:2, 505-51

Deep learning (DL)

  • Fancy subfield that has been applied to all areas of ML, and is the current mainstay of AI research.
  • Describes the architecture of the algorithm.
  • Performance scales with data, e.g. if you give a DL algorithm 10,000 cystoscopy images it will get super accurate, whereas a non-DL algorithm might top out and not get any better past a few hundred images.
  • Focus on stuff above and below before diving in to DL.

Applications in clinical imaging

  • Diabetic retinopathy: the Google paper you’ve seen
  • Cardiac MRI: Arterys, a medical imaging technology company, recently partnered with GE Healthcare’s cardiac MRI technology to better visualize and quantify blood flow inside the heart and dramatically speed up image processing.
  • Dermatology: from Stanford and Google reports the use of deep learning analysis of dermoscope images to distinguish 1) keratinocyte carcinomas versus benign seborrheic keratoses; and 2) malignant melanomas versus benign nevi.
  • Glioma: paper attached; work done by Emory researchers I know (one of whom now chairs Northwestern pathology) reports prediction of overall survival of patients diagnosed with brain tumors from microscopic images of tissue biopsies and genomic biomarkers.
  • Consolidation in chest X-rays

Resources to learn the basics

  • Python 3 (do not learn Python 2)